cs 207 d - computer architecturecomputerarchitecture.yolasite.com/resources/lecture 9reduced.pdf ·...

Computer and Information Sciences College /

Computer Science Department

CS 207 D

Computer Architecture

Lecture 9: Multiprocessors

Challenges of Parallel Processing

First challenge is % of program inherently

sequential

• Second challenge is long latency to

remote memory

1. Application parallelism primarily via new algorithms that have better parallel performance

2. Long remote latency impact both by architect and by the programmer

Forcing Challenges of Parallel Processing

• Long remote latency impact

For example, reduce frequency of remote accesses either by

1) Caching shared data (HW)

2) Restructuring the data layout to make more accesses local (SW)

More Challenges:

– Parallel Programming

– Complex Algorithms

– Memory issues

– Communication Costs

– Load Balancing

Shared‐Memory Architectures

From multiple boards on a shared bus to multiple processors inside a single chip

Caches both

Private data are used by a single processor

Shared data are used by multiple processors

Caching shared data

reduces latency to shared data, memory bandwidth for shared data, and interconnect bandwidth

cache coherence problem

Amdahl’s Law

Architecture design is very bottleneck-driven

1. make the common case fast

2. do not waste resources on a component that has little impact on

overall performance/power

• Amdahl’s Law:

– performance improvements through an enhancement is limited

by the fraction of time the enhancement comes into play

Execution Time After Improvement =

Execution Time Unaffected + Execution Time Affected

Amount of Improvement

Speedup = Performance after improvement

Performance before improvement

Flynn Classification of parallel processing

architectures

Flynn classification of parallel processing architectures

SISD, SIMD, MISD, MIMD

Basic multiprocessor architectures

SISD

SIMD

MISD

MIMD

GMSV

GMMP

DMSV

DMMP

Single data stream

Multiple data streams

Sin

gle

instr

str

eam

M

ultip

le in

str

str

eam

s

Flynn’s categories

Jo

hn

son

’s e

xpa

nsi

on

Shared variables

Message passing

Glo

bal

me

mory

D

istr

ibute

d

me

mory

Uniprocessors

Rarely used

Array or vector processors

Multiproc’s or multicomputers

Shared-memory multiprocessors

Rarely used

Distributed shared memory

Distrib-memory multicomputers

Flynn’s Taxonomy

• Flynn classified by data and control streams in 1966

Single Instruction Single Data

(SISD) or (Uniprocessor)

Single Instruction Multiple Data

SIMD (single PC: Vector, CM-2)

Multiple Instruction Single Data

(MISD) or (ASIP)

Multiple Instruction Multiple Data

MIMD (Clusters, SMP servers)

Computer Architectures

SISD SIMD MIMD MISD

SISD

No instruction parallelism

No data parallelism

• SISD processing architecture example

─ a personal computer processing instructions

and data on single processor

Chapter 17&18

SIMD Multiple data streams in parallel with a single instruction stream

At one time, one instruction operates on many data

Data parallel architecture

Vector architecture has similar characteristics, but achieve the

parallelism with pipelining.

Array processors ─ a graphic processor processing instructions for translation or rotation or other

operations are done on multiple data

• An array or matrix is also processed in SIMD

Chapter 17&18

MISD

Multiple instruction streams in parallel operating on single

Data stream

• Processing architecture example─ processing for critical

controls of missiles where single data stream processed on

different processors to handle faults if any during processing

MIMD

Multiple processing streams in parallel processing

on parallel data streams

MIMD processing architecture example is super

computer or distributed computing

systems with distributed or single shared memory

Chapter 17&18

Flynn Classification

SISD: traditional sequential architecture

SIMD: processor arrays, vector processor Parallel computing on a budget – reduced control unit cost

Many early supercomputers

MIMD: most general purpose parallel computer

today

Clusters, MPP, data centers

MISD: not a general purpose architecture.

Johnson’s expansion

Lecture 1 : Introduction 15

classification of Parallel Processor

Architectures

What is Parallel Architecture

A parallel computer is a collection of processing elements

that cooperate to solve large problems

– Most important new element: it is all about communication !!

• What does the programmer (or OS or compiler writer) think

about?

1) Models of computation

• Sequential consistency?

2) Resource allocation

• What mechanisms must be in hardware

– A high performance processor (SIMD, or Vector Processor)

3) Data access, Communication, and Synchronization

Chapter 17&18

Multiprocessor Basics “A parallel computer is a collection of processing elements

that cooperate and communicate to solve large problems fast.”

• Parallel Architecture = Computer Architecture + Communication

Architecture

• 2 classes of multiprocessors WRT memory:

1. Centralized Memory Multiprocessor

• <few dozen cores

• Small enough to share single, centralized memory with Uniform Memory

Access latency (UMA)

2. Physically Distributed‐Memory Multiprocessor

• Larger number chips and cores than the first.

• BW demands Memory distributed among processors with

Non‐Uniform Memory Access/latency (NUMA)

Multicore: Mainstream Multiprocessors

Multicore chips • IBM Power5

•Two 2+GHz PowerPC cores Shared 1.5 MB L2, L3 tags • AMD Quad Phenom

Four 2+ GHz cores • Per-core 512KB L2 cache

Shared 2MB L3 cache • Intel Core 2 Quad Four cores, shared 4 MB L2

Two 4MB L2 caches • Sun Niagara 8 cores, each 4-way threaded Shared 2MB L2, shared FP • For servers, not desktop All have same shared memory

programming model

Chapter 17&18

Definitions of Threads and Processes

Instruction stream divided into smaller streams (threads)

Parallel execution

Thread in multithreaded processors may or may not be same as software threads

Process: An instance of program running on computer

Thread: dispatchable unit of work within process

Thread switch Switching processor between threads within same process

Typically less costly than process switch

a thread of execution is the smallest sequence

of programmed instructions that can be

managed independently

But First, Uniprocessor Concurrency

•

• Software “thread”:

Independent flow of execution

• Context state: PC, registers

• Threads generally share the same memory space

• “Process” like a thread, but different memory space

• Java has thread support built in, C/C++ supports P-threads library

• Generally, system software (the O.S.) manages threads

• “Thread scheduling”, “context switching”: PC and registers switching

• All threads share the one processor

• Quickly swapping threads gives illusion of concurrent execution

thread is a component of a process. Multiple threads can exist within the

same process but different processes do not share these resources.

Conceptual Diagram

Thread A Thread B Thread C Thread D

Hardware Multithreading (TLP)

Coarse- grained multithreading

Single thread runs until a costly stall

E.g. 2nd level cache miss

Another thread starts during stall for first

Pipeline fill time requires several cycles!

Does not cover short stalls

Less likely to slow execution of a single thread

(smaller latency)

Needs hardware support

PC and register file for each thread

little other hardware

Coarse Multithreading

Stalls for A and C would be longer than indicated in previous slide

Assume long stalls at end of each thread indicated in previous slide

Hardware Multithreading

Fine-grained multithreading

Two or more threads interleave instructions

Round-ribbon fashion

Skip stalled threads

Needs hardware support

Separate PC and register file for each thread

Hardware to control alternating pattern

Naturally hides delays

Data hazards, Cache misses

Pipeline runs with rare stalls

Does not make full use of multi-issue architecture

Fine Multithreading

Skip A

Hardware Multithreading

Simultaneous Multithreading (SMT)

Hyper-threading by Intel

Instructions from multiple threads issued on same cycle

Uses register renaming and dynamic scheduling facility of multi-issue architecture

Needs more hardware support

Register files, PC’s for each thread

Temporary result registers before commit

Support to sort out which threads get results from which instructions

Maximizes utilization of execution units

Simultaneous Multithreading

Skip A

Skip C

Hardware Multithreading Hardware Multithreading (MT)

• Multiple threads dynamically share a single pipeline (caches)

• Replicate thread contexts: PC and register file

• Coarse -grain MT: switch on L2 misses Why?

• Fine grain MT: Does not make full use of multi-issue architecture

• Simultaneous MT: no explicit switching, it converts TLP into ILP

• MT does not improve single-thread performance

Multithreading Categories

Basic Principles

Systems provided of reconfigurable logic are often called

Reconfigurable Instruction Set Processors (RISP). The

reconfigurable logic includes a set of programmable

processing units, which can be reconfigured in the field to

implement logic operations or functions, and programmable

interconnections between them.

Reconfiguration Steps

To execute a program taking advantage of the

reconfigurable logic, use the following Steps.

1. Code Analysis: the first thing to do is to identify

parts of the code that can be transformed for

execution on the reconfigurable logic.

The goal of this step is to find the best trade off

considering performance and available resources

Basic steps in a reconfigurable system

Lecture 1 : Introduction 35

Reconfiguration Steps

2- Code transformation: Once the best candidate parts of

code to be accelerated are found,.

3- Reconfiguration After code transformation, it is time to send

to the reconfigurable system.

4. Input Context Loading: To perform a given reconfigurable

operation, a set of input operands are fetched

5. Execution: After the reconfigurable unit is set and the

proper input operands are fetched

6. Write back: The results of the reconfigurable operation are

saved back to the register file,

Underlying Execution Mechanism

Why is reconfigurable system interesting?

1. Some applications are poorly suited to microprocessor.

2. VLSI “explosion” provides increasing resources.

3. Hardware/Software co-design

4. Relatively new research area.

5. Relatively the same computation

Programming multiprocessor systems

What is concurrency? ● What is a sequential program?

A single thread of control that executes one instruction and when it

is finished execute the next logical instruction

● What is a concurrent program?

A collection of autonomous sequential threads, executing (logically)

in parallel

● The implementation (i.e. execution) of a collection of threads

can be:

1. Multiprogramming – Threads multiplex their executions on a

single processor

2. Multiprocessing – Threads multiplex their executions on a

multiprocessor or a multicore system

3. Distributed Processing – Processes multiplex their executions

on several different machines

Why use Concurrent Programming?

1) Natural Application Structure: The world is not sequential!

Easier to program multiple independent and concurrent

activities.

2) Increased application throughput and responsiveness

3) Not blocking the entire application due to blocking IO

4) Performance from multiprocessor/multicore hardware

5) Parallel execution

6) Distributed systems

7) Single application on multiple machines

8) Client/server type or peer-to-peer systems

Comparing Uniprocessing to Multiprocessing

A Uniprocessor (UP) system has one processor.

A driver in a UP system executes one thread at any given

time.

A Multiprocessor (MP) system has two or more processors.

A driver in this system can execute multiple threads

concurrently.